Erlangen 2026 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 92: Electronics, Trigger, DAQ IV

T 92.2: Vortrag

Freitag, 20. März 2026, 09:15–09:30, KH 00.023

Track Finding with Graph Neural Networks at the ATLAS Event Filter — •Giulia Fazzino, Sebastian Dittmeier, and André Schöning — Physikalisches Institut, Universität Heidelberg, Germany

The upcoming High Luminosity upgrade of the Large Hadron Collider will increase the number of simultaneous interactions per bunch-crossing in the ATLAS experiment from ⟨µ⟩≈ 56 to ⟨µ⟩≈ 200.

To cope with the computational demands resulting from the corresponding rise in data rate, the Trigger and Data Acquisition System of the experiment will undergo several upgrades. The trigger will consist of a hardware trigger and a software trigger, the Event Filter. In the latter, charged particle tracks in the Inner Tracker (ITk) will be reconstructed for event selection. To reduce the computational resources required for this task, the possibility of using hardware accelerators and new tracking algorithms has been extensively investigated over the last years.

One promising approach uses Graph Neural Networks (GNNs) for track finding, one of the most computationally expensive steps of track reconstruction. The algorithm first builds a graph from the hits in the ITk, then uses a GNN to score its edges, and lastly applies a segmentation procedure to generate track candidates. The high parallelizability of the method makes it suited for implementation on FPGAs or GPUs.

This talk will present an overview of GNN-based track finding for the ATLAS Event Filter, with a focus on its implementation and optimization for FPGA deployment.

Keywords: Track Reconstruction; Graph Neural Networks; ATLAS Event Filter